Research map / Multi-agent systems and games / LLM agents

Probe Before You Edit: Probing-Guided Molecular Optimization for LLM Agents in Structure-Based Drug Design

NeurIPS 2026 · Multi-agent systems and games › LLM agents · approach: Robust and efficient agent execution · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes PROBE, a probe-before-edit framework for LLM-agent structure-based drug design that builds a pocket-specific site map and an EditManual from controlled probe edits, then runs a multi-agent affinity/druggability co-optimization loop, reaching SOTA on CrossDocked2020.

Problem

LLM agents for structure-based drug design refine ligands without knowing how the pocket-ligand complex responds to local edits, so single edits rarely improve binding affinity (Vina) and druggability (QED, SA) jointly and often trade one off against the other.

Approach

Before iterative optimization, run a 3x4 grid of controlled probe edits (three strategies at three intensities plus a counterfactual), distill the responses into a site map and a site-level EditManual of allowed/forbidden actions, and use these to guide affinity, druggability and co-optimization agents.

Key results

• MolCRAFT + PROBE reaches Vina -9.203 and success 63.76% on CrossDocked2020 versus 31.40% for MolCRAFT + CIDD, ranking 2nd overall in the table.
• MolJO + PROBE achieves the best overall rank (1) with Vina -9.626 and success 67.67%.
• In the diagnostics, PROBE gets 52.8% joint-improvement steps versus at most 28.4% for baselines.

Surprising findings

• Asking the LLM to write an EditManual from prior knowledge alone gives only small and inconsistent gains; the gain comes from probing evidence.

Limitations

• Relies on Vina docking as the affinity proxy and QED+SA as druggability proxies.
• Probing adds 12 extra Vina calls per molecule, increasing compute relative to 5-round baselines.
• Results rely on GPT-4o-mini as the backbone for all agent roles.

Experimental setup

Datasets: CrossDocked2020

Benchmarks: CrossDocked2020

Metrics: Vina QED SA Success rate MRR QikProp Lipinski LogP MW Rank

Baselines: MOLLM CIDD CIDD+MOO LIDDIA AR Pocket2Mol TamGen LMLF ELILLM IDOLpro TAGMol DrugGPS

Builds on: CIDD MOLLM LIDDIA BRICS PLIP

Closest papers by meaning

• DecompOpt: Controllable and Decomposed Diffusion Models for Structure-based Molecular Optimization (ICLR 2024 · similarity 0.86)
• Enhancing Ligand Validity and Affinity in Structure-Based Drug Design with Multi-Reward Optimization (ICML 2025 · similarity 0.85)
• Generating Novel Leads for Drug Discovery Using LLMs with Logical Feedback (AAAI 2024 · similarity 0.85)
• CBGBench: Fill in the Blank of Protein-Molecule Complex Binding Graph (ICLR 2025 · similarity 0.85)
• Efficient Evolutionary Search Over Chemical Space with Large Language Models (ICLR 2025 · similarity 0.85)
• pCoMole: Pareto-Constrained Molecule Editing with Discrete Flows (NeurIPS 2026 · similarity 0.85)